WEAK CONVERGENCE OF PROBABILITY MEASURES REVISITED

Cabriella saltnetti' Roger J-B we t s e

April 1987 W P-87-30

This research was supported in part by MPI, Projects: "Calcolo Stocastico e Sistemi Dinamici Statistici" and "Modelli Probabilistic" 1984, and by the National Science Foundation.

Working Papers are interim reports on work of the lnternational Institute for Applied Systems Analysis and have received only limited review. Views or opinions expressed herein do not necessarily represent those of the Institute or of its National Member Organizations.

INTERNATIONAL INSTITUTE FOR APPLIED SYSTEMS ANALYSlS A-2361 Laxenburg, Austria

FOREWORD

The modeling of stochastic processes is a fundamental tool in the study of models in- volving uncertainty, a major topic at SDS.

A number of classical convergence results (and extensions) for probability measures are derived by relying on new tools that are particularly useful in stochastic optimization and extremal statistics.

Alexander B. Kurzhanski Chairman

System and Decision Sciences Program

ABSTRACT

The hypo-convergence of upper semicontinuous functions provides a natural frame- work for the study of the convergence of probability measures. This approach also yields some further characterizations of weak convergence and tightness.

CONTENTS

1 About Continuity and Measurability

2 Convergence of Sets and Semicontinuous Functions

3 SC-Measures and SGPrerneasures on 7 (E)

4 Hypo-Limits of SGMeasures and Tightness

5 Tightness and Equi-Semicontinuity

Acknowledgment

Appendix A

Appendix B

References

- vii -

WEAK CONVERGENCE OF PROBABILITY MEASURES REVISITED

Gabriella salinettil and Roger J-B wetsL

luniversiti "La Sapienza", 00185 Roma, Italy. 2~niversity of California, Davis, CA 95616.

1. ABOUT CONTINUITY AND MEASURABILITY

A probabilistic structure - a space of possible events, a sigma-field of (observable)

subcollections of events, and a probability measure defined on this sigma-field - does not

have a built-in topological structure. This is the source of many technical difficulties in

the development of Probability Theory, in particular in the theory of stochastic processes.

Much progress was made, in reconciling the measuretheoretic and topological viewpoints,

by the study of limits in terms of the weak*-convergence of probability measures, also

called weak convergence [5],[16]. In this paper, we approach these questions from a funda-

mentally different point of view, although eventually we show that weak*-convergence

and convergence in the sense introduced here, coincide for probability measures defined on

separable metric spaces. We proceed by a "direct" construction: it is shown that the

spaces of probability measures is in one-bone correspondence with a certain space of

upper semicontinuous functions, called sc-measures, for which there is a natural topology,

and thus an associated notion of convergence. This means that instead of relying on the

pre-dual to generate the notion of convergence, we use the "topological" properties of the

space of probability sc-measures itself, and much insight is gained by doing so.

The major tool is the theory of epi- or hypo-convergence that has been developed in

Optimization Theory to study the limits of (infinitevalued) semicontinuous functions.

Functions are said to hypcxonverge if their hypographs converge (as sets); the hypograph

of a (extended-)real valued function consists of all points on and below its graph. This

"global" view of functions provided by the hypographical approach is particularly appeal-

ing when dealing with limit theorems in Probability Theory. Hypconvergence is not

blinded by what happens at single points, a cause of major chagrin when working with

standard pointwise convergence, but takes into account the behavior of the (converging)

functions in the neighborhood of each single point.

We associate to every probability measure, its restriction to the hyperspace of a

given class of closed sets. It is easy, but crucial, to observe that the resulting function,

called an sc-measure, is upper semicontinuous with respect a certain topology on this

(hyper)space of closed sets. Now, limits can be defined in terms of the hypo-limits of these

sc-measures as is done in Section 4. The structural compactness of the space of upper sem-

icontinuous functions with the hypo-topology is the key to a number of limit results, in

particular in the study of the role played by tightness, cf. Sections 4 and 5. In Section 3,

we show that when the probability measures are defined on the Bore1 sigma-field of a

metric space, hypo-limits of probability sc-measures and weak*-limits of the associated se-

quences of probability measures coincide, providing us with an indirect proof of a Riesz-

type representation theorem for the space of sc-measures. We think that this new charac-

terization of weak*-convergence, that supplements those that have already been studied

extensively [13] and which a t least conceptually is non-standard, is by its intrinsic nature

closely related to the usual notion of convergence of distribution function of random vec-

tors. In fact, one may feel that it would be much more natural to approach convergence in

distribution for random vectors in terms of the hypo-convergence of their distribution

functions rather than through the equivalent but less meaningful notion of pointwise con-

vergence on the continuity set. The reader can convince himself of this, all what is needed

is the definition of hypo-convergence (9, Section :I] and the accompanying geometric

cappretation.

We work in general with probability measures defined on a separable metric space,

but we reserve a special place to the case when in addition the underlying space is com-

pact. For this, there are some technical and didactic reasons - namely in the compact case

the notion of convergence of sets that we introduce corresponds to the usual notion of

set-convergence - but in addition the compact case played an important role in our

research on the convergence of stochastic infima (extremal processes) that originally

motivated this work. This was first elaborated in the presentation of these results in a

communication to the European Meeting of Statisticians a t Palermo (September 1982),

and also a t the Third Course in Optimization Theory and Related Fields a t Erice (Sep

tember 1984). The connections to problems in stochastic homogenization and related

questions in the Calculus of Variations was brought to our attention by a recent article of

De Giorgi [8], whose research is following a path that in some ways, is parallel to ours.

To assume that the domain of definition of the probability measures is compact, or

even locally compact, is not a standard assumption in probability theory, it would ex-

clude a large number of functional spaces that are of capest in the study of stochastic

processes. However that is only true if we restrict ourselves to a "classical" functional

view of stochastic processes. Instead, if we approach the theory of stochastic processes as

in [18 (Sections 3 or 6), 191 where paths are viewed as elements of a space of semicontinu-

ous functions which we equip with the epi-topology, or when paths are identified with

their graphs and the space of graphs is given the topology of set-convergence, then the

underlying space is actually compact. Specific examples of such constructions can be

found in [18, 20) and in the work of Dal Maso and Modica [7] on the stochastic homogeni-

zation problem (the space of integral functionals on Lf,(Rn) is given the epi-topology).

Because set-convergence - or the variant that we introduce here to handle the

infinite dimensional case, is not yet a familiar tool in probabilistic circles - in the appen-

dices the proofs are given in painstaking detail. The alerted reader can of course skip

much or all of this, and devote her or his full attention to the actual substance of the

results.

2. CONVERGENCE OF SETS AND SEMICONTINUOUS FUNCTIONS

Let (E,d) be a separable metric space with r the topology generated by the metric d.

By 7, or 7 (E) , we denote the hyperspace of r-closed subsets of E. We endow 7 with the

topology T that corresponds to the following notion of convergence: for {F; Fn , n E N)

F = T - lim Fn n-a3

if for all z E E ,

d(z, F ) = lim d(z, Fn) ; n-a3

i-e. T is generated by the pointwise convergence of the distance functions {d(', Fn) ,

n E N ) to d(', F). For any nonempty set D c E ,

and

Francaviglia, Lechicki and Levi have extensively studied the properties of T and related

topologies and uniformities [lo], [14]t In particular, they point out that T-convergence can

+They call T the Wijaman topology, but that doea not aeem to be totally appropriate, aince already Choquet in his 1947 paper 'Convergences' (Annalea de 1'Univ. de Grenoble, 23) introduces this notion for aet- convergence.

be characterized in the following terms. For any q > 0, and nonempty set D in E , the

open r)- fattening of D is the set

with q D : = c l a m D the r)- fattening of D. By definition, for some fixed z in E , usually 0 if

E is a linear space, we set

We also reserve the notations

B0,(z) : qo {z), and B,(z) = q{z)

for the open and closed balls of radius q and center z.

PROPOSITION 2.1 [9, Propositions 2.1 and 2.21 Suppose {F; Fn, n E N ) c 7. Then

F = T- lim Fn if and only if n+oo

(i) for all z E E, to every pair 0 < c < q, there corresponds n' such that for all n > n',

(ii) for all z E E, to every pair 0 < c < q, there corresponds n' such that for all n 2 n',

It follows immediately from this proposition that T- lim Fn = @ if and only if to any n+oo

bounded set Q, there corresponds n' such that for all n 5 n',

if d is a bounded metric, this means that the Fn are empty for n sufficiently large.

This notion of convergence for closed sets is related to, but is more restrictive than,

the more standard definition of set-convergence, by which one means that

where

F = lim Fn n+oo

lim sup Fnc F c lirn inf Fn n+oo n+oo

r - cluster points of all seq~ences{z")~=~ I zn E Fn n-tm I

and

lirn inf Fn : = r - limit points of all s e q ~ e n c e s { z " ) ~ ~ I zn E Fn n--roo I I

here U is the Frtkhet filter on N = {1,2, ...) and U# its grill, i.e.

U# = { H E N ( HnH'f @ for all H 'E U ) .

The grill U# consists of all infinite (countable) subsets of N. Since U c U#, we always

have that

lirn inf Fn c lirn sup Fn, n-rm n-tm

and thus F = lirn Fn actually means that equality holds in (2.5). From these definitions,

we immediately have the following proposition; the details can be found in Appendix A.

PROPOSITION 2.2 [lo, Proposition 2.3, Theorem 2.61 Consider {Fn, n=l,,,) a se-

quence in ?(E). Then F = 7- lim Fn implies that F = lim Fn. The converse also holds n-tm n-tm

if (E,d) is boundedly compact (i.e., every closed ball is compact).

As a simple example of a case when the converse does not hold, take E = 11, Fn = {

the unit vector on the n-th axis ). Then lirn Fn = @ whereas 7-lim Fn does not exist.

Some further properties of 7 and set-convergence are reviewed in Appendix A.

Francaviglia, Lechicki and Levy prove that (7, ?) is metrizable and separable [lo,

Theorem 4.61 which of course also implies that it is regular and has a countable base,

since singletons are closed. It is these latter properties that play a key role in what fol-

lows. (One can also rely on the separability of (E, r), and the characterization of conver-

gence provided by Proposition 2.1 to obtain a countable base.) For easy reference we

record these facts in the next theorem. The remaining assertion is well documented in the

literature [ l l ; 15; 9, Proposition 3.21.

THEOREM 2.3 Suppose (E, d) is a separable metric space. The topological space

(7(E), 7) has a countable base and is regular (which also means that it is separable and

metrizable, since it is TI). Moreover it is compact, if in addition (E, d) is locally compact.

We note that there is no loss of generality in introducing 7-convergence in terms of

sequences instead of nets (or filters).

Now let us consider SCU(* 0, I]) , the space of 7-upper semicontinuous functions (u.

sc.) on 7 with values in [0, 11. Recall that a function v(F) - (0, 11 is 7 u. sc. at F when-

ever

v (F ) 2 lim sup v (Fn) n+oo

for all sequences {Fn)r=l with 7-limn., Fn = F.

Next, we endow SCu(* [0, I]) with a convergence structure based on the sequential

convergence of the hypographs. Recall that the hypograph of v : 7 - [0, 11 is the set

hypo v : = {(F, a ) E ( 7 x R) 1 v(F) 2 a) , (2.11)

i.e. all points that lie on or below the graph of v.

DEFINITION 2.4 A sequence of functions {v,, n E N ) in SCU(E 10, 11) hypo-converges

t o v ~ S C ~ a t F ~ 7 i f

(i) whenever F = 7- lim FnJ then n+m

lim sup vn(Fn) 5 v(F) , n+oo

and

(ii) for some sequence {Fn)F'Ll unth F = T- lim FnJ n+ca

lim sup vn(Fn) 2 v(F) . n+oo

Note that the first condition could equivalently be formulated as follows: for any subse-

quence H E FI#, and collection {Fn, n E H) with F = 7-limnE R Fn , (2.12) holds when n

goes to w on H. This observation is also useful in showing [9, Proposition 1.91 that h y p e

convergence (at all F in 7) corresponds to the (sequential) set-convergence of the h y p e

graphs, i.e.,

v = hypelim vn if and only if hypo u = lim hypo vn . (2.14) n+oo

If E is a compact metric space, hypeconvergence of upper semicontinuous real valued

functions can be characterized in terms of convergence of the Hausdorff distances between

the graphs [4].

Of crucial importance to the ensuing development is the following compactness

result. A very elegant proof appears in [2, Theorem 2-22] that relies directly on the

definition of hypeconvergence. It can also be derived from a general theorem about set-

convergence, for the convenience of the reader, a proof can be found in Appendix B.

THEOREM 2.5 Suppose (E , d ) is a separable metric space, and 3 is the hyperspace of

closed subsets of E, that we equip with the topology 7 of pointunse convergence of the dis-

tance functions. Then (SCU(R [0, I ] , hypo) is sequentially compact, i.e., any sequence

{v, E SCU(J; [0, j.]), n E N } contains a subsequence that hypo-converges to a junction v

in SCU(J; [0, I]). If, i n addition (E , d ) is locally compact, then (SCU(E [0, I]), hypo) is a

regular compact topological space.

3. SC-MEASURES AND SC-PREMEASURES O N ?(E)

An sc-premeasure X is a (set-) function on 3(E) with the following properties:

i. nonnegativity: X(F) > 0 for all FEZ . . 11. increasing (n~ndecreasin~): if for any F1, F2 in ?,

X(Fl) 5 X(F2) whenever Fl c Fa;

. . . 111. X is 7-u.sc. (upper semicontinuous) on ?.

It is finitely sub-additive if for an F1, Fa, in 3,

X(F1) + X(F2) I X(F,uF,) + X(F1nF,) -

If, it actually is finitely additive, i.e. for any F1, F2 in ?,

then X is an sc-measure. It is a probability sc-measure if in addition

Sc-measures (semicontinuous measures) defined on ?(E) are of course intimately re-

lated to measures defined on B(E), the Borel field generated by ?, the T-closed subsets of

E . Conceptually, however, there is a basic difference between measures and sc-measures.

The measure-calculus relies on the underlying sigma-field structure (countable additivity,

etc.), the calculus of sc-measures is topological in nature, and consequently provides a

richer structure for studying convergence, and other limit questions. It is, however, possi-

ble to identify probability measures and sc-measures as we show next.

THEOREM 3.1 There i s a one-to-one correspondence between probability measures on

B(E) and probability sc-measures on 3(E). More precisely, given a probability measure p

on B(E), then the restriction of p to 3 (E) i s a probability sc-measure. And given A, a pro-

bability sc-measure, there ezists a unique probability measure p on B(E) such that p = X

on 3( E) .

PROOF Suppose p is a probability measure on B and X is its restriction to 3. Clearly,

it is enough to check if X is 7-u.sc. Since Thas a countable base, it suffices to show that

lim sup X(Fn) 5 X(F) n+m

whenever F = 7-lim Fn. First observe that

F = lim sup Fn = n cl (U Fn) 1 n u p"'= F' n+m HEN nEH HEN nEH

where the first equalities follow from Proposition 2.2, and (2.7). Since p is a probability

measure and X(Fn) = p(Fn) we have that

If X is a probability sc-measure on 3, we set p = X on 3 and define for every open set

G, p(G) = 1-X(E\G). We see that p is an increasing finitely additive set-function on A,

the field consisting of finite cups of open and closed sets. We can now appeal to the stan-

dard argument to extend p to a probability measure on B , the sigma-field generated by A

[ I , Theorem 1.3.10). This extension is unique. In fact, since (E,T) is a separable metric

space, for every A E B, we have that p(A) = sup {X(F) I FC A ) .

As can be expected from the preceding theorem, measures and sc-measures have

many common properties. Of immediate capest are certain continuity properties used in

the sequel.

By bdy D we denote the T-boundary of a set D c E.

PROPOSITION 3.2 Suppose A: 3 - [0,1] is a sc-premeasure. Then given any nonemp-

t y F E 3, the function

has at most countably many discontinuity poinh and X(F) = lim,loX(cF). Thus, if X is a

probability sc-measure, the family of sets {cF I c 2 0) contains at most countably many

sets such that X(bdy c F ) > 0.

PROOF First note that F = T-lim,lo c F, see A.5. Since X is 1-u.sc a t F and increas-

ing, and hence X(F) 5 lim infClo X(c F) , it follows that X(F) = limElo X(c F). The asser-

tion of a t most countable discontinuity points follows directly from the (topological) argu-

ment given in [6, iv p.41 that applies to all monotone functions. Finally, if X is a probabil-

ity sc-measure, for any c1 > 0 with X(bdy c1 F ) > 0, we have for any c1 < €1

Taking the above into account, it yields

which shows that c1 is a discontinuity point of c -X(cF) and there are a t most a count-

able number of such points.

The correspondence between probability measures and sc-measures carries over to

the natural convergence notions for probability measures (weak convergence) and sc-

measures (hypo-convergence). In those terms it is a "bicontinuous" correspondence, as is

demonstrated next. Recall that weak convergence, or more precisely weak* convergence,

of a sequence of probability measures {p,: B - [0,1], nE N ) to a probability measure

p:B(E)-[0,1], denoted p = weak*- lim-pn means that n+m

for any bounded continuous g: E - R, or equivalently [5, Theorem 2.11

lim inf pn(G) > p(G) for all open sets G, n+m

lim pn(A) = p(A) for all sets A E cont p, n+w

where

cont p:= {A E B I p(bdy A ) = 0)

THEOREM 3.3 Suppose (E, r) i s a separable metric space, and {p; pn, nE N) is a

family of probability measures on B ( E ) , and {A; An, nE N) the corresponding family of

probability sc-measures on J (E) . Then

whenever

X = hypo-lim An. n-rm

This is also a necessary condition if (E,d) is a Polish space, i.e. (E,d) is also complete.

PROOF If X = hypo-lim A n then (3.5) follows directly from (2.12), since pn = A n on 7, n-rm

and hence p = weak*-lim p,.

To prove the converse we start with (2.13), i.e. we show that given any F E J there

exist a sequence with F = 7-lim Fn such that n-rm

X(F) I lim inf Xn(Fn). n-m

We take F nonempty, since otherwise the inequality is trivially satisfied. From (3.6) it

follows that for every e > 0,

p(eo F ) 2 lim inf pn(eeF) n-rm

and thus to every 6 > 0 there corresponds n6 such that for all n 2 n6:

In particular, this means that to every k E N we can associate nk, with nk+' > nk, such

that for all n 2 nk

With Fn = k - ' ~ , 6, = k-' whenever n E [nk,nk+,], by A.5 we have that F = 7-lim Fn

and

X(F) 5 lim inf [Xn(Fn) + 6,] = lim inf Xn(Fn). n-m n-m

There remains to show that (2.12) holds when (E,d) is a Polish space. Since (E,d) is

complete p is tight, i.e. given any 6 > 0 there exist a compact set K such that

p(K) > 1-6. This implies that for any e > 0, p(c1 (E\eK)) I 6. For such a compact set

K, r sufficiently small, and for any sequence with F = T-lim Fn, we have

lim sup X,(F") 5 lim sup pn(FnnrK) + lim sup pn(cl (E\rK)) n+m n+m n+m

and by A.8, (3.5) and the above, this yields

lim sup Xn(Fn) 5 lirn sup pn(r'F) + 6 5 p(r'F) + 6 = X(r'F) + 6 n+m n+m

where r' > 2r. Since by A.5, T-limE,lo r'F = F, X is T-u-sc. (Theorem 3.1 and the

definition of sc-premeasure) and 6 is arbitrary, it follows that

lirn sup Xn(Fn) 5 X(F), n+m

which completes the proof.

Theorem 3.3 can be viewed as giving a new characterization of weak convergence for

probability measures. The classical results of Prohorov [5, Section 61 could be obtained as

a direct consequence of this characterization, and the compactness results of Section 2.

However it is more enlightening to derive it as a consequence of the properties of the

hypo-limits of sc-measures as is done in the next section.

4. HYPO-LIMITS OF SC-MEASURES AND TIGHTNESS

In this section we are capested in the properties of the limit functions of a sequence

of probability sc-measures. In view of Theorem 2.5 we know that there always will be a

limit function, a t least for some subsequence, what cannot be guaranteed is that this limit

function is a probability sc-measure. We being with a general result about sc-premeasures.

LEMMA 4.1 The space of sc-premeasures is sequentially compact with respect to hypo-

convergence. In particular, if {A,, nE N ) is a sequence of sc-premeasures such that

Xn(E) = 1 for all n, and X:=hypo-lim An, then X is a sc-premeasure with X(E) = 1. n+m

PROOF Every sequence of sc-premeasurea has a hypo-convergent subsequence

(Theorem 2.5). The limit function is then T-u-sc., and clearly it is nonnegative. We need

to show that it is an increasing function. Suppose X = hypo-lim An, and let us consider n+m

any pair Fl c Fg in 7. Since X is the hypo-limit of the An, it follows from (2.13) and

(2.12) that there exists a sequence with F1 = T-lim Fr such that X(FI) = lim Xn(FY).

Since F2 = T- lirn Fa u F r , see A.3, it follows that v+m

X(F2) 2 lim sup Xn(F2 u Fr) 2 lim sup Xn(FT) = X(F1). n+m n+m

Finally, X(E) = 1 whenever Xn(E) = 1 for all n, since by (2.12)

X(E) 2 lim sup Xn(E) = 1 , n+m

and X(E) 5 lim inf Xn(Fn) 5 1 whatever be the sequence {Fn, nE N ) n+m

T-converging to E .

Thus the hyp l i rn i t X of a sequence of probability sc-measures is a sc-premeasure.

But not much more can be said except for a super-additivity property, a t least not

without making some further assumption about (E,T); it is easy to verify that X always

satisfies:

X(Fl) + X(F2) 2 X(F1uF2) for all F1,F2 in J .

In general, X is neither finitely additive, nor is A(@) = 0. A s the ensuing development will

show these two properties are not unrelated. Let us begin by giving a necessary and

sufficient condition for having A(@) = 0.

LEMMA 4.2 Suppose {A,, nE N ) is a sequence o j probability sc-measures on J (E) ,

and X = hypo-lim An. Then A(@) = 0 ij and only ij to every c > 0, there corresponds a n+m

closed ball C, and n, such that for all n 2 n,,

PROOF Fix any c > 0. Proposition 2.1 and the definition of hypeconvergence yield the

existence of a sequence {Fn, nE N ) such that for all n n,,

FnnC, = @, and A(@) = lim Xn(Fn). n+m

Since for all n 2 n,, Xn(Fn) 5 1 - Xn(C,) < c, it implies that 0 5 A(@) < c. This holds

for all c > 0, which means that A(@) = 0.

Let us now prove the converse. We argue by contradiction. Suppose A(@) = 0, but

for some c > 0 and every closed ball C, there exists HC E )I# such that for all

nEHC, Xn(C) 5 1- c. Since

it follows that for all nE Hc, Xn(cl (E\C)) > c. The fact that X is the hypelimit of the

A n implies that for all closed ball C,

c 5 lim sup Xn(cl (E\C)) 5 X(c1 (E\C)). n+m

Consider any increasing sequence {Cn, nE N) of closed balls that T-converges to E. This

means that T-lim cl (E\ Cn) = 2, cf. A.6. The preceding inequality and the T-u.sc. of X n+m

(Lemma 4.1) would yield the following contradiction.

0 < c 5 lim sup X(c1 (E\Cn) 5 X(d) = 0. n+m

This leads us to the following observations. A collection of probability sc-measures A

on 7 (resp. measures M on 8) is said to be tight, if to every c > 0 there corresponds a

compact set K, c E such that for all, but a finite number, of X in A (resp. p E M)

X(K,) > 1 - c, (resp. p(K,) > 1 - €1. (4-2)

If the metric space E has compact closed balls, we can rewrite the assertion of Lemma 4.2

as follows: A(@) = 0 if and only if the sequence {A,, nE N) is tight. But, as it turns out,

having A(@) = 0 is all what is needed to obtain the finite additivity of X when E is locally

compact. To show this one first proves that in the locally compact case, the hypo-limit of

a sequence of probability sc-measures is always a finitely sub-additive sc-premeasure.

Next, it is shown that if in addition this hypo-limit has A(@) = 0, then X is actually finite-

ly additive and hence X is a probability sc-measure. This means: if the {A,, nE N) are

probability sc-measures on 7(E) with E a locally compact separable metrizable space, and

X = hypo-lim An, then X i s a probability sc-measure if and only if the sequence n+m

{A,, nE N) is tight. This follows from the fact that E locally compact, separable and

metrizable admits a boundedly compact metric 1231. When E is given this metric, we are

in the following situation: T-convergence and set convergence coincide, and

(SCU(5 [0, I]), hypo) is compact; details can be found in [19].

The same assertion can be made if E is simply Polish - and this is proved below -

but tightness plays then a double role. As in the locally compact case, it guarantees that

there is no escape of the probability mass "at infinity", but it is also used to generate "re-

lative compactness" in the space of probability measures. Tightness already enters in the

proof that the hypelimit is sub-additive. It essentially allows us to restrict our attention

to compact subsets of E where T-limits coincide with the standard set-limits, and up to

some technical adjustments we can rely, as in the locally compact case, on the built-in re-

lative compactness of the space of u-sc. functions (with the hypetopology).

THEOREM 4.3 Suppose {A,, nE N) i s a sequence of probability sc-measures on ?(E)

where E is a separable metric space, and X = hypo-lim An. Then n+w

(i) if the sequence {A,, nE N) is tight, X is a probability sc-measure;

(ii) if i s a probability sc-measure and E i s a Polish space, the sequence

{An nE N) i s tight.

PROOF If the {A,, nE N) are tight, then X is a sc-premeasure with X(E) = 1 (Lemma

4.1), but also A(@) = 0 (Lemma 4.2). To show that X is sub-additive (3.1), observe that

since X is the hypelimit of the An, for any pair F1,F2 in 7, there exists sequences such

that 7-lim Fr = F1 and 7- lim F; = F2 such that n+a,

lim Xn(Fr) = X(Fl), and lim Xn(F,") = F2. n+oo n+oo

Given c > 0, let Kc be a compact set such that for all, but finitely many,

n: Xn(K,) > 1 - c. Since the A n are probability sc-measures:

Taking lim sup on both sides, using the fact that X is the h y p l i m i t of the A n (2.12),

since F l U F 2 = 7-lim (FTuF,") by A.3, and F1 n F2 n K, = 7-lim n+oo

(cf. Proposition 2.2, A.2 and A.4), we obtain

X(Fl) + X(F2) 5 X ( F p F2) + X(F,n F2n K,) + 6

where for the last inequality we used the fact that X is increasing since i t is sc-

premeasure. This yields (3.1) since c > 0 is arbitrary. To complete the proof of part (i),

there remains only to show that

Since A(@) = 0, we may as well assume that F1 and F2 are nonempty, the inequality being

trivially satisfied otherwise. Let {Rn, nE N ) and {Sn, nE N ) be sequences of sets such

that 7- lim R n = F,uF,, 7- lim S n = F l n F2, and n+m n4oo

X(F,U F,) = lim Xn(Rn) , X ( F l n F2) = lirn Xn(Sn). n-+a, n+m

The existence of such sequences follows as before from (2.13) and (2.12). Pick any q > 0 .

With F' = cl (E\q F,), and using the fact that the An are probability sc-measures, we

have

Xn(Rn) + Xn(Sn) 5 Xn(Rnnq F,) + Xn(Rnn F') +

Moreover, the sequence {A,, nE N ) is tight. Let K , be a compact set such that

Xn(K,) > 1 - (€12) for n sufficiently large. Thus for n sufficiently large

Taking lim sup on both sides, using the fact that X is the hypo-limit of the {A,, n€ N) ,

and since

7 - lim [ ( ( R n n F') u S n u F1) n K,] = F , n K ,

and

18 = 7 - lirn ( R n n S n n F' n K,) n-03

as follows from Proposition 2.2, A.2, A.3 and A.4, we have

We obtain (4.3) from the above by observing that X is 7-U.SC a t F,, 7- lim qF2=F2 by n-03

A.5, and that c > 0 can be chosen arbitrarily small.

To prove the converse, part (ii), observe that since (E , r ) is a Polish space every pro-

bability measure [5 , Theorem 1.41, and thus every probability sc-measure (Theorem 3.1)

is tight. In particular, this means that for all c > 0 there exists K c E , compact such that

X ( K ) > 1 - C . In turn, this implies that for all q > 0 , X(E\q0K) 2 e . Since

A = hypo-lim A,,

lim sup A,(E\ q0K) 5 A(E\q0K) 5 r n+oo

from which it follows that there exists n , such that for all n _> n , and q > 0:

Now, we use the fact that q > 0 is arbitrary, that by A.5, K = T-lim qK and that the A,

are T-u.sc. at K to conclude that for all n 2 n,, A,(K) > 1 - 3r. This means that the se-

quence {A,, n€ N ) is tight, since there is such a compact set K for any r > 0. n The theorems of Prohorov and Varadajaran [5, Section 6; 15, Theorem 6.71 are im-

mediate consequences of the above. As in [5] or 1161, we say that a family M of probabili-

ty measures is relatively compact (with respect to the weak* topology on M) if any se-

quence {p,, n€ N ) c M contains a subsequence weak converging to a probability meas-

ure.

COROLLARY 4.4 Prohorov's Theorems Suppose (E,d) i s a separable metric space and

M i s a family of probability measures on B(E). Then, if M i s tight i t i s relatively compact.

Moreover if (E,d) is a Polish space, then relative compactness of M implies tightness.

PROOF Let A = {A: 7(E) - [0,1] I A = p on F, p E M) be the associated family of

probability sc-measures, cf. Theorem 3.1. Lf M is tight, so is A. Moreover, any sequence in

A contains a hyp+convergent subsequence (Theorem 2.5) which is necessarily tight, and

hence its hypelimit is a probability sc-measure (Theorem 4.3. (i)). The first assertion

now follows directly from Theorem 3.3.

For the converse, let us first consider the case when M = {p,, nE N ) . It follows

directly from the definition of relative compactness, Theorem 3.3, Theorem 4.3. (ii), the

definition of tightness and the fact every probability sc-measure on the Polish space (E,d)

is complete, that for all r > 0 there exists K, such that not only p,(K,) > 1 - r for all

n E N , but also p(K,) > 1 - r where p is any (probability) measure in the weak* closure

of M . To complete the proof simply observe that the weak* topology (on the space of

measures on 8) is separable, and thus there exists a dense subset {p,, nE N ) C M

whose weak* closure is (uniformly) tight.

We note that the separability of weak* topology can be obtained as a consequence of

Theorems 3.1 and 2.3.

5. TIGHTNESS AND EQUI-SEMICONTINUITY

Theorem 2.5 can be viewed as a generalization of a version of Helly's Theorem for

probability sc-measures. The standard formulation of Helly's Theorem is, however, in

terms of a "pointwise" convergence of the distribution. From the results of this section it

will follow that hypo-convergence of %-measures can be given a pointwise characteriza-

tion, which also leads us to relate tightness to an equi-upper semicontinuity condition "at

infinity".

The relationship between pointwise and hypo-convergence has been studied in (91.

Neither one implies the other, unless the collection of functions is equi-u.sc. [9, Theorem

2.91. A family U c SC,(7; [0,1]) is equi-upper semicontinuous at F (with respect t o the

7-topology) if to every c > 0 there corresponds a 7-neighborhood V of F such that for all

UE U,

sup v ( V ) < v ( F ) + c . VEV

PROPOSITION 5.1 Suppose {A; A,, nE N ) are probability sc-measures such that

X = hypo-lim A,. Then ,--roo

(i) the sequence i s equi-u.sc. on cont A,

(ii) @ E cont X

where cont X = {F E 3 1 X (bdyF) = 0).

PROOF From Theorem 3.3, (3.7) and Theorem 3.1, it follows that for all F E cont A,

lim X,(F) = X(F). This means that on cont X we have both hypo- and pointwise con- ,--roo

vergence, and this only occurs if the sequence is equi-u.sc. on cont X 19, Theorem 2.181; a

direct proof is given in Appendix B.

Part (ii) follows from: bdy @ = @ and A(@) = 0.

THEOREM 5.2 Suppose {A; A,, n€ N ) are probability sc-measures. Then

X = hypo-lim A, if and only if X(F) = lim X,(F) for all F E cont A. n-oo ,--roo

PROOF For the "only if" part, see above. For the converse we rely on Proposition 3.2

and the fact that X is 7-u.sc. Indeed they imply that given any cl > 0, there always exists

c E (O,cl) such that c F E cont X and X(c F ) < X(F) + cl. Because of pointwise conver-

gence on cont A, we have that

lim sup X,(F) 5 lim X,(c F ) = X(c F ) < X(F) + cl, n--rcx, n h o o

whatever be c1 > 0. Hence lim sup A,(F) 5 A(F) for all F E 3. But this via Theorem 3.1, n4m

(3.5)) and Theorem 3.4 yields the hypconvergence of the A, to A.

If we are given a sequence {A,, nE N ) of probability sc-measures that hypo-

converges to A, and it also pointwise converges on cont A, this would not yet imply that A

is a probability sc-measure: finite additivity and A(@) = 0 might still fail to be satisfied. If

(E,r) is boundedly compact however, it does suffice to have @ E cont A, as follows from

the next proposition.

PROPOSITION 5.3 Suppose (E,r) is a boundedly compact separable metric space, and

{A,, nE N) are probability sc-measures on 3(E). Then the following statements are

equivalent:

(i) the sequence {A,, nE N) is tight;

(ii) the sequence {A,, nE N) is equi - u-sc. at @;

(iii) for any sequence {Fn, nE N ) &th @ = lim Fn, we have n--+m

lim sup An(Fn) = 0. n--+m

PROOF Using the characterization of basic neighborhood systems of @, and recalling

that the closed balls are compact (E,r) are compact, we see that the {A,, nE N ) are

equi - u.sc. at @ if and only if to every c > 0, there corresponds a compact ball K, such

that for all n

X,(F)LA,(@)+c=c whenever F n K , = @ ,

or equivalently A,(K,) > 1 - 6 . In other words, if and only if the sequence is tight.

The equivalence of (i) and (iii) follows from Theorem 4.3 and Proposition 5.1.

Proposition 5.3 and Theorem 5.2 allow us, in the compact case, to rephrase Theorem

4.3 as follows: Suppose (E,r) is a compact metric space, {A,, nE N ) is a sequence of pro-

bability sc-measures on 3(E) such that A = lim A, on cont A. Then A is a probability n+m

sc-measure if and only if the {A,, nE N ) are equi - u.sc. at @.

Acknowledgment We are very grateful to Professor G. Beer for a substantial number

of suggestions and pointers, in particular in connection with the boundedly compact as-

sumption in Proposition 2.2, and its relationship to local compactness.

APPENDIX A

We begin by a collection of facts about set- and 7-convergence that are used in the

text. We always assume that (E,d) is a separable metric space.

A . l 7-convergence implies set-convergence

PROOF To being with we show that if F = 7-lim Fn, then F c lim inf Fn. There is n+m

nothing to prove if F = @, so let us assume that F is nonempty. By Proposition 2.1, if

z E F and c > 0, then Bo,(z) nFn # @ for all n sufficiently large. This means that for all

HEH#, x E T-cl(un Fn) and hence, by (2.9), z E lirn inf Fn.

Next, let us show that F > F': = lim sup Fn, or equivalently that E \ F c E\FI. n+m

There is nothing to prove if F = E, so let us assume E \ F # @. Suppose z E E\F, then

there exists 9 > 0 such that FnBub~(z ) = @. By Proposition 2.1 this implies that for any

c ~ ( 0 , T), for n sufficiently large FnnB,(z) = @. Thus there exists H E U such that

z r - c l ( ~ , ~ ~ F " ) , a n d b y (2.7)thisimpliesthatz l i m s u p F n , i . e . , z E E \ F ~ .

The fact that 7-convergence and set convergence are the same when (E,d) is a Eu-

clidean space is well know, see [17, Theorem 2.21 for example. Beer [4] pointed out that

this equivalence can only be obtained for spaces whose closed balls are compact.

A.2 Fl=lim Fr, F2=lim F," implies FlnF2 = lim((FlnF2)u(F,"nFz")1

PROOF Set F = F1 n F2, Fn = Frn F,". Since F c FuFn, F c lirn inf(Fu Fn). If

z E lirn sup(FuFn) this means that there exists HEN#, Z"E FuFn for all n E H such

that z = lirn zn. If Z"E F infinitely often then z E F, otherwise

z E lim sup Fn c lirn sup Frnlim sup F," = FlnF2 = F; the inclusion can be proved from

the definition (2.6) of lirn sup, or one can consult [12, $25. 111. Hence lirn sup(Fu Fn) c F

and, with the above, this yields (2.5).

A.9 F1= 7-lim F r , Fz= 7-lim Ft implies FluF2= 7 - l i m ( F r ~ F t ) .

PROOF Use the definition of 7-convergence

A.4 F = lim Fn, Fc Fn. Then F n K = 7-lim Fnn K

where K is any compact set in E.

PROOF We use the characterization of 7-convergence given by Proposition 2.1. For

any pair 0 < r < q, F n K nBO,(z) # @ implies that Fn n K n BO,,(z) # @ for all n, since

F c Fn. Now if for some q > 0, F n K n B,(z) = @, but there exists r > 0, and h E U #

such that Fn n K n B,(z) = @ for all n E N , it would mean that

limsupn,, (Fn n K n B,(z)) is nonempty, contradicting the assumption that

F n (K n B,(z)) = @ since

lirn sup(Fn n K n B,(z)) c (lim sup Fn) n (K n B,(z)) c F n ( K n B,(z)) . n-oo n-oo

PROOF Since (3, 7) is separable (Theorem 2.3) we only need to consider this limit in

terms of a sequence {c,, n~ N ) with lim rn = 0. If F = @ then the result is a direct

consequence of the definition of r-fattening of the empty set and the comment that follows

Proposition 2.1. Now suppose F is nonempty, then for all q > 0 such that FnBO,,(z) # @ it follows that r,FnB,,(z) # @ for all n, since F c rn F. If FnB,,(z) = @ then for any

r E(O q), for n sufficiently large rn FnB,(z) = @ as follows from the fact that E is normal.

A.6 {Cn,nE N ) is an increasing sequence o j closed balls unth E=7-limCn.

Then {cl(E\Cn),n€ N ) is decreasing and @ = 7-lim cl(E\Cn).

PROOF The sequence is clearly decreasing. Suppose D is any bounded set, then for n

sufficiently large D c int Cn and hence D n cl(E\Cn) = @. Since this holds for any

bounded set D, from Proposition 2.1 it follows that {cl (E\Cn), n E N ) 7-converges to @.

A.7 F = 7-lim Fn, or more generally F = lim Fn, and K c E compact.

Then for all c>O, there eztsts n, such that for all nzn,, F n n K c c F .

PROOF If F = 6, then F n n K = 6 for n sufficiently large (Proposition 2.1), in which

case the inclusion is obviously satisfied. If F is nonempty and F n n K is not included in cF,

i t means that there exists H E U# such that for all ~ E H , FnnKn(E\cF) # 6. Passing to

a subsequence if necessary, it means that there exists {zn, n E H) such that

Z"E FnnKn(E\cF) , and by (2.6) we have

lim zn = z E (lim sup Fn)nKncl (E\cF) = Fncl (E\cF)nK

But this latter set is empty, contradicting the possibility that Fnn K is not included in

c F for n sufficiently large.

A . 8 F = 7-lim Fn, 6 # K compact and c > 0. Then for n suficiently large

Fn n c K c c tF where ct > 2c.

PROOF The case F = 6 is argued as in A.7. Otherwise, for contradiction purposes,

suppose that there exists H E U# such that for all n in H, dist (zn, Fn \ ctF) < c for some

zn E K. Passing to a subsequence, if necessary, we have that lim zn = z E K n cF, since

K is compact, c F = 7-lim c Fn and every zn E c Fn. On the other hand d(zn, F ) > ct - c

and thus d(z, F ) > c, again by 7-convergence, contradicting the possibility that z E c F.

APPENDIX B

B.l Proof ojTheorem 2.5

This theorem, an application of a general result about convergence of sets to the

space of hypographs, has a long history that starts with a result of Zoretti [22] in the

complex plane; Hausdorff, Lubben, Urysohn, Blaschke and Marczewski, having contribut-

ed in bringing the theorem in its present form. We give a direct proof patterned after the

argument used in (121 for sequences of sets (see also [21] for sequences of functions).

To being with let us observe that to any sequence {An€ SCU(7; [0, 11). nE N ) we

can associate an upper hypo-limit hypo-1s A n defined by

(hypo-Is Xn)(F):= lim sup Xn(Fn) s u P ~ ~ U # s u p { ~ n , n E ~ I F=T-limFn) nEH

and a lower hypo-limit hypeli A n defined by

(hypo-li Xn)(F):= lim inf Xn(Fn). s u p { ~ n , n ~ ~ I F=T-limFY) n-m

As a direct consequence of the definition of hypelimit, we have that X = hypo-lim A n if

and only if

If we denote by U(F), the T-neighborhood system of F , the upper and lower hypo-limit

can also be expressed as

hypo-Is Xn(F) = inf lim sup sup AEU(F) n-oo F'EA

Xn(F1).

and

hypo-li Xn(F) = inf lim inf sup AEU(F) n-oo F'EA Xn(F').

Let {A1,l=l, ...) be a countable open base for 7 , see Theorem 2.3. Note that for each

1, the sequence

has a t least one cluster point in (the compact space) R. Let N ~ E X # determine a subse-

quence such that

lim (sup An) ezists. nEN1 A 1

Define recursively Nl c Nl-l such that for all 1

lirn (sup An) exists. "€4 A1

By diagonalization, construct N' c N as follows

N': ={nl 1 nl is the 1-th member of Nl) .

Since for all l,{nl, 1 = 1, ...) c Nl, we have that for all 1:

lirn ~ E N ~ ~ ~ F E ~ ~

Xn(F) exists.

Now for any F E 7, we have

= inf lim sup sup ff~.Af(F) n€N1 F'EA

Xn(F')

= inf lim SUP A[E.A~(F) nEN' F'EA[ Xn(F')

= inf lim inf sup AE.A~(F) n€N1 F'EA Xn(F1)

= (hypo-li Xn)(F). n€ N'

Since this holds for any F E ~ , we have that (B.1) is satisfied with

X = hypo-1s A n = hypo-li A n for the subsequence N' c N .

The second assertion follows from the fact that T-convergence coincides with the

standard set-convergence (Proposition 2.2) and the second part of Theorem 2.3, if one ob-

serves that the hyperspace of hypographs is a closed subset of the hyperspace of closed

subsets of 7(E) X R where 7 is 7-compact since E is locally compact; for details see [9,

Corollary 4.21.

B.2 Direct proof of Proposition 5.1. (i).

Arguing by contradiction, let us assume that the {A,, nE N ) are not equi-upper

semicontinuous a t F E cont A. This means, there exists c > 0 such that to every T-

neighborhood V of F there corresponds Nv EN# such that for all n E Nv

Now, let {Vk, k=l , ...) a (countable) fundamental neighborhood system of F , the existence

of such a system follows from Theorem 2.3. For every k, let Hk€ H # be such that (B.2)

holds with U = Uk. Pick nkE Hk\{nlr...rnk-l), and choose F "~ E Vk such that

Let N' = {nk, k=l, ...) and define the collection {Fn, nEN) as follows:

Fn:= F~~ if n E N' and n = nk.

then F = T-lim Fn, and

lim sup Xn(Fn) 2 lim sup Xnt(~") k+m

2 e + lim sup Xnt(F) = e + X(F) k-+m

where the last equality follows from Theorems 3.1 and 3.3, and (3.7) since F E cont A.

But this is in contradiction with the hypcxonvergence of the A n to A, in particular with

REFERENCES

[ I ] Ash, R.: Real Analysis and Probability, Academic Press, New York, 1972.

[2] Attouch, H.: Variational Convergence for Functions and Operators. Pitman, Boston 1985.

[3] Beer, G.: Metric spaces with nice closed balls and distance functions for closed sets, Bull. Australian Mathematical Society, 35(1987), 81-96.

Beer, G.: Upper semicontinuous function and the Stone approximation theorem, J. Approzimation Theory, 34, 1982, p. 1-11.

Billingsley, P.: Convergence of Probability Measures, J. Wiley & Sons, New York, 1968.

Chung, K.: A Courae in Probability Theory, Harcourt, Brace and World, Inc., New York, 1968.

Dal Maso, G. and L. Modica: Nonlinear stochastic homogenization, Technical re- port, Universitk Pisa, 1984.

De Giorgi, E.: On a definition of I'-convergence of measures, in Multifunctions and Integrands: Stochastic Analysis, Approzimation and Optimization, ed. G . Salinetti, Springer Verlag Lecture Notes in Mathematics 1091, Berlin, 1984, pp 150-159.

Dolecki, S., G. Salinetti and R. Wets: Convergence of functions: Equi- semicontinuity, Trans. Amer. Math. Soc., 276 (1983), 409-429.

Francaviglia, A. Lechicki and S. Levi: Quasi-uniformization of hyperspaces and con- vergence of nets of sernicontinuous multifunctions, J. Mathematical Analysis and Ap- plications 112(1985), 347-370.

Frolik, Z.: Concerning topological convergence of sets, Czechoslovak Math. J , 10 (1960), 168-180.

Kuratowski, C.: Topologae I, PNW, Warzawa, 1958.

[13] Kurtz, T.: Approaches to weak convergence, in Multifunctiona and Integrands: Sto- chastic Analysis, Approzimation and Optimization, ed. Salinetti, Springer Verlag Lecture Notes in Mathematics 1091, Berlin, 1984, pp 173-183.

[14] Lechicki, A. and S. Levi: Wijsman convergence in the hyperspace of a metric space, Manuscript Universita di Pisa, 1986.

[15] Lowen-Colebunders E.: On the convergence of closed and compact sets, Pacific J. Math., CVIII (1983), 133-140.

[16] Parthasaraty K.: Probability Measures on Metric Spaces Academic Press, New York, 1967.

1171 Salinetti, G. and R. Wets: On the convergence of closed-valu'ed measurable multi- functions, Transactions American Mathematical Society, 266(1981), 275-289.

(18) Salinetti, G. and R. Wets: Convergence in distribution of measurable multifunctions (random sets), normal integrands, stochastic processes and stochastic infima, Mathematics of Operations Research, 11(1986), 385-419.

1191 Salinetti, G. and R. Wets: On the hypo-convergence of probability measures, Optim- ization and Related Fields, eds. R. Conti, E. De Giorgi, and F. Gianessi, Springer Verlag Lecture Notes in Mathematics 1190, Berlin, 1986, pp 371-396.

[20] Salinetti, G. and R. Wets: Random semicontinuous functions, WP-86-47, Interna- tional Institute for Applied Systems Analysis, 1986.

[21] Wets R.: On a compactness theorem for epi-convergent sequences of functions, Mathematical Programming, eds. R. Cottle, M. Kelmanson, and B. Korte, Elsevier (North Holland), 1984, 347-355.

(221 Zoretti, L.: Sur les fonctions analytiques uniformes qui posskdent un ensemble par- fait discontinu de points singuliers, J. Math. Pures Appl., 1 (1905), 1-51.